Accuracy-Relevance trade-off in Transfer Learning for Object Detection
Dishant Parikh, Saurabh Sachdev · 2020 International Conference for Emerging Technology (INCET) · 2020
This paper addresses the problem of accuracy while using different datasets to train the parent model while using transfer learning. We show how the training time and accuracy of the custom object detector depend on the parent model's dataset. We show comprehensive empirical evidence that there is a strong dependence of datasets, while there is a possibility to train any model up to the desired accuracy if trained for a longer time, or if the relevance with the custom dataset is increased. The more the relevance, the faster the training and more is the accuracy of the custom detector. The difference in accuracy, given the same time to train, but different parent dataset is 10-12%. When using Ava as parent dataset we obtain 68.4%. By merely changing the parent dataset to COCO, we achieved a 20% increase in the mAP score.